Umetrics Suite Blog

How to predict the future and optimize your production process

November 30, 2017

Using real-time data analytics monitoring has become the accepted way to monitor processes in several industries. The goal is to detect and diagnose issues as they happen, which is a great leap forward compared to traditional analysis conducted in retrospect. This has been highlighted in a previous blog post.

Is predicting the future magic - or is it data analytics

However, even if you use this type of highly advanced real-time analytics, you can only analyze what has happened up to the current time point. You don’t know yet if problems will occur further on in the process. The ideal would of course be to be able to predict the future of the process and use that information to avoid future problems from happening.

If we know and understand the future, we can do changes now to improve future results. Can we predict where the process is likely to go, based on the way it has run so far? What will, for example, the final yields or final qualities look like? Can we detect and mitigate deviations before they happen? Can we get an optimized result?

Trajectory of a process-1.png

The image shows the trajectory of a process (the black line). The green line shows where the process preferably should continue, according to the production model, but the future trajectory is unknown.

How to predict the future

Predicting the future may sound like magic but there are methods to predict the future of a process. A simplified example is a situation where there is a correlation between two process variables, x1 and x2, as shown in the image below.

Correlated variables.png

The image shows two correlated variables, x1 and x2.

If we only measure x1, what are the likely future values of x2? If we plot our estimates onto a scatter plot we can draw a line of best fit, and then use that line to estimate the likely results of x2 for new measurements of x1 (see image below). In this way, we can fill in the missing values that are unknown. The method to get these values is called imputation.


The image shows two correlated variables. Variable x1 is measured and variable x2 is estimated from the correlation between x1 and x2, i.e., the missing value of x2 is imputed.

In real-world applications, similar estimations of future values of critical process parameters are done using sophisticated missing value imputation algorithms that involve multiple variables at the same time.

In many cases, the beginning of a process correlates with later performance. If we use the data that we have collected up to the current time point, we can use imputation to predict where the process is likely to go. And when we can predict the future of the process, we can also intervene and optimize it.

If you are interested in the mathematics behind it, a PCA or PLS score space model is used like a calibration curve of the process. Based on the data that you have measured you can read off the missing data from the calibration curve.

Predictive Monitoring.png

The image shows the trajectory of a process. The black line shows the result of the process so far. The dark blue line shows the predicted “open loop” future performance of the rest of the process.

Model Predictive Control: How to use predictions to optimize performance

Model Predictive Control is a methodology where you use predictions to control and optimize a process. The goal is to find the best future settings of your process variables, in order for the future of the process to be as you want it to be. For example, if you run a biological process you might collect data for variables such as pH and temperature. Model Predictive Control can then be used to adjust the future values of the pH and temperature to optimize the future of the process.

An outstanding tool to do that is a module inside SIMCA-online called Control Advisor.

Control Advisor: Get more value from your data

If you are familiar with the Umetrics™ Suite of Data Analytics Solutions you probably know that SIMCA is a tool that you can use to build multivariate production models. Once you are satisfied with your model, you can run your model in your manufacturing environment in real-time using SIMCA-online.

A great feature inside SIMCA-online is the module called Control Advisor. Control Advisor is a supervisory type of control tool that advises you how you should run your process to avoid problems before they arise and to get optimal final results. With Control Advisor you will simply get more value from the data that you already have.

If you are interested in the mathematics, the Control Advisor uses a method called imputation by regression, IBR, a patented imputation method that utilizes PLS regression.

How Control Advisor works in practice

Control Advisor can be used in both continuous processes and batch type processes. Control Advisor uses existing measurements of the process (up till current time point), future known setpoints and the multivariate model to propose “smart” adjustments to critical process parameters. In the image below, an example trajectory is shown following such a proposed adjustment from Control Advisor.

Predictive Monitoring 2.png

The image shows the trajectory of a process. The black line shows the result of the process so far. The dark blue line shows the predicted “open loop” future performance of the rest of the process. The bright blue line shows the predicted “closed loop” future performance following a process intervention as suggested by Control Advisor.

The image below shows an overview of a process in SIMCA-online using the Control Advisor, which suggests how variables could be changed to get optimal future process performance.

Control Advisor in Simca-online Umetrics Suite

Want to know more?

To find out more about Control Advisor in SIMCA-online, and how it can help optimize your processes and business results, book a free demonstration.

Book a free demo


Have a question or comment?
Leave it below.


Topics: Real Time Data Analytics, Manufacturing Quality Control, Manufacturing Processes

Jonas Elfving

Written by Jonas Elfving

Product Manager at Sartorius Stedim Data Analytics